log_put.c

这是国外的resip协议栈

	DB_LSN *lsnp;
	u_int32_t logfile;
{
	DB_CIPHER *db_cipher;
	DB_ENV *dbenv;
	DB_LSN lsn;
	DBT t;
	HDR hdr;
	LOG *lp;
	int need_free, ret;
	u_int32_t lastoff;
	size_t tsize;
	u_int8_t *tmp;

	dbenv = dblp->dbenv;
	lp = dblp->reginfo.primary;

	DB_ASSERT(logfile == 0 || logfile > lp->lsn.file);

	/* If we're not at the beginning of a file already, start a new one. */
	if (lp->lsn.offset != 0) {
		/*
		 * Flush the log so this file is out and can be closed.  We
		 * cannot release the region lock here because we need to
		 * protect the end of the file while we switch.  In
		 * particular, a thread with a smaller record than ours
		 * could detect that there is space in the log. Even
		 * blocking that event by declaring the file full would
		 * require all threads to wait here so that the lsn.file
		 * can be moved ahead after the flush completes.  This
		 * probably can be changed if we had an lsn for the
		 * previous file and one for the current, but it does not
		 * seem like this would get much more throughput, if any.
		 */
		if ((ret = __log_flush_int(dblp, NULL, 0)) != 0)
			return (ret);

		/*
		 * Save the last known offset from the previous file, we'll
		 * need it to initialize the persistent header information.
		 */
		lastoff = lp->lsn.offset;

		/* Point the current LSN to the new file. */
		++lp->lsn.file;
		lp->lsn.offset = 0;

		/* Reset the file write offset. */
		lp->w_off = 0;
	} else
		lastoff = 0;

	/*
	 * Replication may require we reset the log file name space entirely.
	 * In that case we also force a file switch so that replication can
	 * clean up old files.
	 */
	if (logfile != 0) {
		lp->lsn.file = logfile;
		lp->lsn.offset = 0;
		if ((ret = __log_newfh(dblp, 1)) != 0)
			return (ret);
	}

	DB_ASSERT(lp->db_log_inmemory || lp->b_off == 0);
	if (lp->db_log_inmemory &&
	    (ret = __log_inmem_newfile(dblp, lp->lsn.file)) != 0)
		return (ret);

	/*
	 * Insert persistent information as the first record in every file.
	 * Note that the previous length is wrong for the very first record
	 * of the log, but that's okay, we check for it during retrieval.
	 */
	memset(&t, 0, sizeof(t));
	memset(&hdr, 0, sizeof(HDR));

	need_free = 0;
	tsize = sizeof(LOGP);
	db_cipher = dbenv->crypto_handle;
	if (CRYPTO_ON(dbenv))
		tsize += db_cipher->adj_size(tsize);
	if ((ret = __os_calloc(dbenv, 1, tsize, &tmp)) != 0)
		return (ret);
	lp->persist.log_size = lp->log_size = lp->log_nsize;
	memcpy(tmp, &lp->persist, sizeof(LOGP));
	t.data = tmp;
	t.size = (u_int32_t)tsize;
	need_free = 1;

	if ((ret =
	    __log_encrypt_record(dbenv, &t, &hdr, (u_int32_t)tsize)) != 0)
		goto err;
	__db_chksum(t.data, t.size,
	    (CRYPTO_ON(dbenv)) ? db_cipher->mac_key : NULL, hdr.chksum);
	lsn = lp->lsn;
	if ((ret = __log_putr(dblp, &lsn,
	    &t, lastoff == 0 ? 0 : lastoff - lp->len, &hdr)) != 0)
		goto err;

	/* Update the LSN information returned to the caller. */
	if (lsnp != NULL)
		*lsnp = lp->lsn;

err:	if (need_free)
		__os_free(dbenv, tmp);
	return (ret);
}

/*
 * __log_putr --
 *	Actually put a record into the log.
 */
static int
__log_putr(dblp, lsn, dbt, prev, h)
	DB_LOG *dblp;
	DB_LSN *lsn;
	const DBT *dbt;
	u_int32_t prev;
	HDR *h;
{
	DB_CIPHER *db_cipher;
	DB_ENV *dbenv;
	DB_LSN f_lsn;
	LOG *lp;
	HDR tmp, *hdr;
	int ret, t_ret;
	size_t b_off, nr;
	u_int32_t w_off;

	dbenv = dblp->dbenv;
	lp = dblp->reginfo.primary;

	/*
	 * If we weren't given a header, use a local one.
	 */
	db_cipher = dbenv->crypto_handle;
	if (h == NULL) {
		hdr = &tmp;
		memset(hdr, 0, sizeof(HDR));
		if (CRYPTO_ON(dbenv))
			hdr->size = HDR_CRYPTO_SZ;
		else
			hdr->size = HDR_NORMAL_SZ;
	} else
		hdr = h;

	/* Save our position in case we fail. */
	b_off = lp->b_off;
	w_off = lp->w_off;
	f_lsn = lp->f_lsn;

	/*
	 * Initialize the header.  If we just switched files, lsn.offset will
	 * be 0, and what we really want is the offset of the previous record
	 * in the previous file.  Fortunately, prev holds the value we want.
	 */
	hdr->prev = prev;
	hdr->len = (u_int32_t)hdr->size + dbt->size;

	/*
	 * If we were passed in a nonzero checksum, our caller calculated
	 * the checksum before acquiring the log mutex, as an optimization.
	 *
	 * If our caller calculated a real checksum of 0, we'll needlessly
	 * recalculate it.  C'est la vie;  there's no out-of-bounds value
	 * here.
	 */
	if (hdr->chksum[0] == 0)
		__db_chksum(dbt->data, dbt->size,
		    (CRYPTO_ON(dbenv)) ? db_cipher->mac_key : NULL,
		    hdr->chksum);

	if (lp->db_log_inmemory && (ret = __log_inmem_chkspace(dblp,
	    (u_int32_t)hdr->size + dbt->size)) != 0)
		goto err;

	if ((ret = __log_fill(dblp, lsn, hdr, (u_int32_t)hdr->size)) != 0)
		goto err;

	if ((ret = __log_fill(dblp, lsn, dbt->data, dbt->size)) != 0)
		goto err;

	lp->len = (u_int32_t)(hdr->size + dbt->size);
	lp->lsn.offset += (u_int32_t)(hdr->size + dbt->size);
	return (0);
err:
	/*
	 * If we wrote more than one buffer before failing, get the
	 * first one back.  The extra buffers will fail the checksums
	 * and be ignored.
	 */
	if (w_off + lp->buffer_size < lp->w_off) {
		DB_ASSERT(!lp->db_log_inmemory);
		if ((t_ret = __os_seek(dbenv,
		    dblp->lfhp, 0, 0, w_off, 0, DB_OS_SEEK_SET)) != 0 ||
		    (t_ret = __os_read(dbenv, dblp->lfhp, dblp->bufp,
		    b_off, &nr)) != 0)
			return (__db_panic(dbenv, t_ret));
		if (nr != b_off) {
			__db_err(dbenv, "Short read while restoring log");
			return (__db_panic(dbenv, EIO));
		}
	}

	/* Reset to where we started. */
	lp->w_off = w_off;
	lp->b_off = b_off;
	lp->f_lsn = f_lsn;

	return (ret);
}

/*
 * __log_flush_pp --
 *	DB_ENV->log_flush pre/post processing.
 *
 * PUBLIC: int __log_flush_pp __P((DB_ENV *, const DB_LSN *));
 */
int
__log_flush_pp(dbenv, lsn)
	DB_ENV *dbenv;
	const DB_LSN *lsn;
{
	int rep_check, ret;

	PANIC_CHECK(dbenv);
	ENV_REQUIRES_CONFIG(dbenv,
	    dbenv->lg_handle, "DB_ENV->log_flush", DB_INIT_LOG);

	rep_check = IS_ENV_REPLICATED(dbenv) ? 1 : 0;
	if (rep_check)
		__env_rep_enter(dbenv);
	ret = __log_flush(dbenv, lsn);
	if (rep_check)
		__env_db_rep_exit(dbenv);

	return (ret);
}

/*
 * __log_flush --
 *	DB_ENV->log_flush
 *
 * PUBLIC: int __log_flush __P((DB_ENV *, const DB_LSN *));
 */
int
__log_flush(dbenv, lsn)
	DB_ENV *dbenv;
	const DB_LSN *lsn;
{
	DB_LOG *dblp;
	int ret;

	dblp = dbenv->lg_handle;
	R_LOCK(dbenv, &dblp->reginfo);
	ret = __log_flush_int(dblp, lsn, 1);
	R_UNLOCK(dbenv, &dblp->reginfo);
	return (ret);
}

/*
 * __log_flush_int --
 *	Write all records less than or equal to the specified LSN; internal
 *	version.
 *
 * PUBLIC: int __log_flush_int __P((DB_LOG *, const DB_LSN *, int));
 */
int
__log_flush_int(dblp, lsnp, release)
	DB_LOG *dblp;
	const DB_LSN *lsnp;
	int release;
{
	struct __db_commit *commit;
	DB_ENV *dbenv;
	DB_LSN flush_lsn, f_lsn;
	DB_MUTEX *flush_mutexp;
	LOG *lp;
	size_t b_off;
	u_int32_t ncommit, w_off;
	int do_flush, first, ret;

	dbenv = dblp->dbenv;
	lp = dblp->reginfo.primary;
	flush_mutexp = R_ADDR(&dblp->reginfo, lp->flush_mutex_off);
	ncommit = 0;
	ret = 0;

	if (lp->db_log_inmemory) {
		lp->s_lsn = lp->lsn;
		++lp->stat.st_scount;
		return (0);
	}

	/*
	 * If no LSN specified, flush the entire log by setting the flush LSN
	 * to the last LSN written in the log.  Otherwise, check that the LSN
	 * isn't a non-existent record for the log.
	 */
	if (lsnp == NULL) {
		flush_lsn.file = lp->lsn.file;
		flush_lsn.offset = lp->lsn.offset - lp->len;
	} else if (lsnp->file > lp->lsn.file ||
	    (lsnp->file == lp->lsn.file &&
	    lsnp->offset > lp->lsn.offset - lp->len)) {
		__db_err(dbenv,
    "DB_ENV->log_flush: LSN of %lu/%lu past current end-of-log of %lu/%lu",
		    (u_long)lsnp->file, (u_long)lsnp->offset,
		    (u_long)lp->lsn.file, (u_long)lp->lsn.offset);
		__db_err(dbenv, "%s %s %s",
		    "Database environment corrupt; the wrong log files may",
		    "have been removed or incompatible database files imported",
		    "from another environment");
		return (__db_panic(dbenv, DB_RUNRECOVERY));
	} else {
		/*
		 * See if we need to wait.  s_lsn is not locked so some
		 * care is needed.  The sync point can only move forward.
		 * The lsnp->file cannot be greater than the s_lsn.file.
		 * If the file we want is in the past we are done.
		 * If the file numbers are the same check the offset.
		 * This all assumes we can read an integer in one
		 * state or the other, not in transition.
		 */
		if (lp->s_lsn.file > lsnp->file)
			return (0);

		if (lp->s_lsn.file == lsnp->file &&
		    lp->s_lsn.offset > lsnp->offset)
			return (0);

		flush_lsn = *lsnp;
	}

	/*
	 * If a flush is in progress and we're allowed to do so, drop
	 * the region lock and block waiting for the next flush.
	 */
	if (release && lp->in_flush != 0) {
		if ((commit = SH_TAILQ_FIRST(
		    &lp->free_commits, __db_commit)) == NULL) {
			if ((ret = __db_shalloc(&dblp->reginfo,
			    sizeof(struct __db_commit),
			    MUTEX_ALIGN, &commit)) != 0)
				goto flush;
			memset(commit, 0, sizeof(*commit));
			if ((ret = __db_mutex_setup(dbenv, &dblp->reginfo,
			    &commit->mutex, MUTEX_SELF_BLOCK |
			    MUTEX_NO_RLOCK)) != 0) {
				__db_shalloc_free(&dblp->reginfo, commit);
				return (ret);
			}
			MUTEX_LOCK(dbenv, &commit->mutex);
		} else
			SH_TAILQ_REMOVE(
			    &lp->free_commits, commit, links, __db_commit);

		lp->ncommit++;

		/*
		 * Flushes may be requested out of LSN order;  be
		 * sure we only move lp->t_lsn forward.
		 */
		if (log_compare(&lp->t_lsn, &flush_lsn) < 0)
			lp->t_lsn = flush_lsn;

		commit->lsn = flush_lsn;
		SH_TAILQ_INSERT_HEAD(
		    &lp->commits, commit, links, __db_commit);
		R_UNLOCK(dbenv, &dblp->reginfo);
		/* Wait here for the in-progress flush to finish. */
		MUTEX_LOCK(dbenv, &commit->mutex);
		R_LOCK(dbenv, &dblp->reginfo);

		lp->ncommit--;
		/*
		 * Grab the flag before freeing the struct to see if
		 * we need to flush the log to commit.  If so,
		 * use the maximal lsn for any committing thread.
		 */
		do_flush = F_ISSET(commit, DB_COMMIT_FLUSH);
		F_CLR(commit, DB_COMMIT_FLUSH);
		SH_TAILQ_INSERT_HEAD(
		    &lp->free_commits, commit, links, __db_commit);
		if (do_flush) {
			lp->in_flush--;
			flush_lsn = lp->t_lsn;
		} else
			return (0);
	}

	/*
	 * Protect flushing with its own mutex so we can release
	 * the region lock except during file switches.
	 */
flush:	MUTEX_LOCK(dbenv, flush_mutexp);

	/*
	 * If the LSN is less than or equal to the last-sync'd LSN, we're done.
	 * Note, the last-sync LSN saved in s_lsn is the LSN of the first byte
	 * after the byte we absolutely know was written to disk, so the test
	 * is <, not <=.
	 */
	if (flush_lsn.file < lp->s_lsn.file ||
	    (flush_lsn.file == lp->s_lsn.file &&
	    flush_lsn.offset < lp->s_lsn.offset)) {
		MUTEX_UNLOCK(dbenv, flush_mutexp);
		goto done;
	}

	/*
	 * We may need to write the current buffer.  We have to write the
	 * current buffer if the flush LSN is greater than or equal to the
	 * buffer's starting LSN.
	 *
	 * Otherwise, it's still possible that this thread may never have
	 * written to this log file.  Acquire a file descriptor if we don't
	 * already have one.
	 */
	if (lp->b_off != 0 && log_compare(&flush_lsn, &lp->f_lsn) >= 0) {
		if ((ret = __log_write(dblp,
		    dblp->bufp, (u_int32_t)lp->b_off)) != 0) {
			MUTEX_UNLOCK(dbenv, flush_mutexp);
			goto done;
		}

		lp->b_off = 0;
	} else if (dblp->lfhp == NULL || dblp->lfname != lp->lsn.file)
		if ((ret = __log_newfh(dblp, 0)) != 0) {
			MUTEX_UNLOCK(dbenv, flush_mutexp);
			goto done;
		}

	/*
	 * We are going to flush, release the region.
	 * First get the current state of the buffer since
	 * another write may come in, but we may not flush it.
	 */
	b_off = lp->b_off;
	w_off = lp->w_off;
	f_lsn = lp->f_lsn;
	lp->in_flush++;
	if (release)
		R_UNLOCK(dbenv, &dblp->reginfo);

	/* Sync all writes to disk. */
	if ((ret = __os_fsync(dbenv, dblp->lfhp)) != 0) {
		MUTEX_UNLOCK(dbenv, flush_mutexp);
		if (release)
			R_LOCK(dbenv, &dblp->reginfo);
		ret = __db_panic(dbenv, ret);
		return (ret);
	}

	/*
	 * Set the last-synced LSN.
	 * This value must be set to the LSN past the last complete
	 * record that has been flushed.  This is at least the first
	 * lsn, f_lsn.  If the buffer is empty, b_off == 0, then
	 * we can move up to write point since the first lsn is not
	 * set for the new buffer.
	 */
	lp->s_lsn = f_lsn;
	if (b_off == 0)
		lp->s_lsn.offset = w_off;

	MUTEX_UNLOCK(dbenv, flush_mutexp);